I have been reading, my conclusions:
1- It seems that only Nvidia GPUs have CUDA,
2- It seems that GPUs that support CUDA are the ones needed in order to train models.
I borrowed x2 Sapphire Radeon Nitro Rx 470 4GB, is there a possibility to use them for training? (is there a flow in my logic)
What you are looking for is the ROCm platform.
As this is not related to OpenNMT, you would need to do some research on how to run TensorFlow or PyTorch on a ROCm-compatible hardware.